Arhitekture umjetnih neuronskih mreža za analizu signala EEG-a

master's thesis

• Title (english): Artificial Neural Network Architectures in EEG Signal Analysis
• Author: Lucija Planinić
• Mentor: Mario Cifrek (mentor)
• Committee member: Mario Cifrek (predsjednik povjerenstva)
• Committee member: Igor Krois (član povjerenstva)
• Committee member: Željka Lučev Vasić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-16, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Elektroencefalografija (EEG) je metoda snimanja električne moždane aktivnosti. Analiza signala EEG-a omogućuje primjenu elektroencefalografije u raznim sustavima, a u tu svrhu se sve više koriste metode strojnog učenja od kojih se ističu umjetne neuronske mreže. U ovom radu korišten je skup signala EEG-a zdravih ispitanika i onih kojima je dijagnosticiran afektivni poremećaj, te su signali klasificirani u dvije klase: „zdravi“ i „bolesni“. Implementirana je konvolucijska neuronska mreža za klasifikaciju signala i autoenkoder za ekstrakciju važnih značajki. Značajke dobivene autoenkoderom zatim su korištene za klasifikaciju pomoću plitkih modela za klasifikaciju (SVM, KNN i Adaboost) i za grupiranje algoritmima za nenadzirano učenje (k-srednjih vrijednosti, aglomerativno grupiranje i DBSCAN). Konvolucijska neuronska mreža i autoenkoder su trenirani tri puta: na nebalansiranom skupu podataka, na skupu balansiranom težinama i na skupu balansiranom poduzorkovanjem. Rezultati su pokazali da je od testiranih modela konvolucijska neuronska mreža najuspješnija u klasifikaciji signala EEG-a (prosječna točnost i F1 mjera u iznosu od 0,997). Balansiranje poduzorkovanjem imalo je bolje rezultate od balansiranja težinama. Grupiranjem je pokazano da se značajke prirodno grupiraju prema dijagnozi ispitanika.

Abstract (english)

Electroencephalography (EEG) is a method of recording the electrical activity of the brain. EEG signal analysis allows the application of electroencephalography in various systems. For this purpose, machine learning methods are increasingly used, of which artificial neural networks stand out. In this paper, a set of EEG signals of healthy subjects and those diagnosed with an affective disorder was used, and the signals were classified into two classes: “healthy” and “not healthy”. A convolutional neural network for signal classification and an autoencoder for feature extraction were implemented. Features obtained by the autoencoder were then used for classification using shallow classification models (SVM, KNN and Adaboost) and for clustering using unsupervised learning algorithms (k-means, agglomerative clustering, and DBSCAN). The convolutional neural network and autoencoder were both trained three times: on an unbalanced dataset, on a dataset balanced using weights and on a dataset balanced by undersampling. The results showed that of the tested models, the convolutional neural network was the most successful in the classification of EEG signals (average accuracy and F1 score of 0,997). Balancing by undersampling had better results than balancing using weights. Clustering showed that features are naturally grouped according to the diagnosis of the subject.

• Keywords: elektroencefalografija (EEG); umjetne neuronske mreže; duboko učenje; konvolucijske neuronske mreže; autoenkoderi
• Keywords (english): electroencephalography (EEG); artificial neural networks (ANN); deep learning; convolutional neural networks (CNN); autoencoders
• Language: croatian
• URN:NBN: urn:nbn:hr:168:564206
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-13 13:30:26